On 19 November HD Moore, chief research officer at Dragos, published the results of a five-month experiment. From July through November his team sent “approximately two million prompts containing unique URLs and canary token credentials to widely used AI services,” then watched what happened to each secret after inference. Each secret went into exactly one prompt to one provider, so any later use of it points back to a single request, a single model, a single route and a single privacy setting.
He reports three kinds of failure. At least three hosting providers had been breached and were having requests copied by an attacker. Resellers in an inference marketplace logged customer prompts and acted on them. And providers fetched or stored prompt contents in ways that contradicted the privacy flags on the request.
The number I keep coming back to is in the third category. One Chinese AI lab’s systems “fetched the secret URLs from more than half of the prompts that we sent to its models between July and October, including roughly two-thirds of the requests marked as ‘no data collection’ or Zero Data Retention.” On those requests the flag didn’t work. If anything, the flagged requests were fetched at a higher rate than the rest.
Where the flag stops
To read that number you need the architecture. An AI router offers one API endpoint and forwards each request to a pool of models, sometimes at different hosting companies, sometimes through another router first. Moore is blunt about what that means: “every hop can see the entire prompt and response.” Privacy controls are set by the customer, at the account, the key or the request, and the router is supposed to forward a ZDR request only to providers that offer the same guarantee.
Moore’s advice to readers contains the sentence that explains most of the results: “a privacy flag only binds the first hop.” The ZDR finding is the clean version. Several U.S. hosting providers resell the lab’s models, and prompts sent to them “triggered the exact same scraper,” so a customer who chose a U.S. provider to keep data in the U.S. still had the prompt read by the lab. In the oddest incident, the secret link from a prompt sent with the ZDR flag to a U.S.-based provider showed up 24 days later in a Feishu chat in China and was opened within two minutes from three desktops and three phones in Beijing and Shanghai.
The breaches show the same thing from the other side. The largest involved “a couple hundred” of the bait AWS keys, sent “within a few short windows in late October,” and “every request we sent during those windows was stolen.” Some came from requests the provider had rejected before any model ran, which places the copying at the entrance. In the other two breaches some of the stolen requests carried the ZDR flag. Once an attacker is inside, the flag doesn’t matter.
I wrote in October about a GLM-5.3 deployment on Cloudflare that returned fragments of other customers’ prompts. That leak was a bug, counted after the fact by the operator. These leaks were deliberate, and they were counted by someone outside who had planted the evidence in advance.
What a canary can and cannot count
Planting unique credentials and waiting for them to be used is an old measurement design, and the literature on it is a useful check on how to read Moore’s percentages.
The closest precedent is Tripwire, presented by DeBlasio, Savage, Voelker and Snoeren at the 2017 Internet Measurement Conference. They registered honey accounts at third-party websites, each tied to an email account with the same password at a major email provider, and treated any login to the email account as evidence that the website had been breached. Across about 2,302 sites with successful registrations they detected 19 compromises. They say outright that the design undercounts. Easy-password accounts were only registered after a hard-password registration seemed to succeed, and “this biases our results to under-report compromises.” On top of that, login data was lost from 20 March to 1 June 2015 because of a misunderstanding about the provider’s retention limits.
The second precedent is closer to an adversarial channel. Peng and colleagues, at AsiaCCS in 2019, leaked the credentials of 150 honey email accounts to live phishing sites, one per site, and monitored them for at least 50 days. Every one of those credentials went to a party whose whole business is collecting credentials. Seven accounts received logins.
That 7 in 150 is the number I’d hold up against the Dragos results. In the Peng study every single credential was known to have reached a criminal, and fewer than one in twenty was used to log in within the monitoring window. A canary fires only when the recipient decides to act on it. Silence means either nobody saw the secret or somebody saw it and didn’t bother. The experiment can’t tell those apart.
That changes how I read the Dragos findings in two ways. First, the lab’s rate is high because its reader was a machine: a scraper that “typically showed up two or three days after the request,” with no person in the loop deciding whether a link looked worth trying. For automated fetching, the firing rate probably sits close to the true exposure rate. For the human-driven cases, like the marketplace seller who opened two secret URLs in a desktop browser a day and a half later and started guessing usernames, or the Mac in France typing SSH passwords by hand, the detections are a floor. The stored copies that nobody happened to act on during the study don’t show up anywhere.
The second point runs the other way, and it matters more for anyone choosing a provider. A route where no canary fired hasn’t been cleared. Moore writes that the bait at the breached providers was “processed along with real customer requests, and those were almost certainly copied too.” The canaries were the only requests in those streams that could report back.
The flag and the policy
The ZDR result has an older relative in web privacy. In the WWW 2018 proceedings, Timothy Libert audited the privacy policies of more than 200,000 websites against the third-party data flows actually observed on them. Fewer than 15% of attributed flows were disclosed. On Do Not Track, the browser flag that was supposed to let users opt out, 7% of policies contained the string “do not track” at all, and in a manual sample, 64.80% of policies that mentioned DNT “explicitly do not honor it.”
DNT and ZDR have the same structure. The customer sets a flag on the request, and the receiving party alone decides whether to honor it, with no way for the customer to see what happened. Libert could at least read the policy that refused the signal. A ZDR customer usually can’t see the downstream hop at all. Dragos had to build what Moore calls “something like traceroute for AI inference,” embedding an image URL that each fetching system would reveal itself through and sending slightly broken requests whose nested error envelopes identify each router. Of about two thousand routes traced, several hundred passed through a system Dragos could not identify. One marketplace seller was found “quietly” adding about a thousand hidden tokens to every prompt.
Who knew
The last precedent concerns disclosure. Of Tripwire’s 19 detected compromises, the authors found only three with any external sign of a breach. Six of the eighteen sites they contacted responded. When the paper was written, none had notified users, although one said it would force a password reset.
Moore’s version of the same pattern: “none of the three providers knew about their breach until we notified them.” The first provider initially attributed the activity to “a misconfiguration of internal tooling, then a problem with partner data processing,” until the investigation found its internal credentials had also been compromised. It “has still not publicly disclosed the incident,” and the stolen data was still driving login attempts in late October. For the late-October intrusions, one provider “has already published a preliminary security notice.” Dragos says it is now working with about a dozen providers.
Tripwire’s authors ended with an open question that fits here: how much of this kind of evidence an operator needs before it will tell its users. A canary tells the operator that it leaked and roughly when. It doesn’t tell it how, which gives an operator that would rather not disclose plenty of room to say it can’t corroborate.
What I’d take from it
Moore’s recommendations are reasonable and short: self-host if you can afford it, go directly to a first-tier provider otherwise, pick upstream providers yourself instead of accepting a router’s default pool, avoid reseller marketplaces, and seed your own canaries. The last one is the one I’d insist on. It costs almost nothing, and it is the only item on the list that produces evidence instead of depending on a promise.
For buyers, the measurement point is the one to keep. “More than half” and “roughly two-thirds” are already alarming, and for the human-run leaks they are probably low, because a canary only counts the readers who acted. A ZDR flag can’t be checked from the customer’s side. A canary can, in one direction only: when it fires, the prompt leaked, and when it stays silent, nothing has been shown either way.
References
- Moore, HD. “Three breaches, endless leaks: Your AI supply chain needs canaries.” Dragos, 19 November 2026. dragos.com/blog/ai-supply-chain-security-canary-tokens
- Goodin, D. “Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.” Ars Technica, 20 November 2026. arstechnica.com
- DeBlasio J, Savage S, Voelker GM, Snoeren AC. “Tripwire: Inferring Internet Site Compromise.” Proceedings of the 2017 Internet Measurement Conference, November 2017. doi:10.1145/3131365.3131391
- Peng P, Xu C, Quinn L, Hu H, Viswanath B, Wang G. “What Happens After You Leak Your Password: Understanding Credential Sharing on Phishing Sites.” Proceedings of the 2019 ACM Asia Conference on Computer and Communications Security, July 2019. doi:10.1145/3321705.3329818
- Libert T. “An Automated Approach to Auditing Disclosure of Third-Party Data Collection in Website Privacy Policies.” Proceedings of the 2018 World Wide Web Conference, 2018. doi:10.1145/3178876.3186087